Healthcare
A U.S. fertility clinic needed to streamline ultrasound follicle monitoring without trading speed for quality. We built an AI-driven detection and measurement system that processes ultrasound videos, tracks follicles across frames, and generates measurements for a doctor to review.
Custom Software Development
Testing & QA
UX/UI Design
Follicle monitoring in ultrasound required a fully manual review and measurement workflow. Doctors had to navigate ultrasound videos frame by frame, identify follicles visually, place calipers to capture maximum horizontal and vertical diameters, and repeat the process for every finding.
The client required a software that could automate follicle detection, segmentation, and tracking across video frames and produce clinically accurate measurements. Based on the requirements, performance had to be optimized for diagnostic reliability and coverage rather than speed, and measurement outputs had to be provided in metric units. The system had a robust pixel-to-millimeter calibration to convert image geometry into accurate diameters, perimeters, and surface areas.
We implemented Mask R-CNN (using ResNet) and extended it with cross-frame tracking, so the system follows the same follicle across multiple frames. Pixel masks are into meaningful units using a calibration module that detects the ultrasound hatch marks (5 mm spacing), computes a pixel-to-mm ratio, and so calculates diameters, perimeter, and surface area per follicle.
Pre- and post-processing removes ultrasound noise and low-contrast boundaries, so a clinician sees only clear results and can track validation metrics. The software supports videos (which are automatically splitted into frames), single images, and folders.
Multi-format ingestion and frame extraction
Accepts ultrasound videos, single images, and folders; automatically extracts frames from video inputs and normalizes them into a consistent processing format.
Applies filtering and refinement steps tailored to noisy, low-contrast ultrasound data to stabilize contours and reduce artifacts before measurement is computed.
Generates precision/recall analytics and run-level summaries to support clinical validation, iterative training, and reproducible QA cycles.
Exports measured findings in a structured, clinician-friendly format (per study/per frame/per detected instance), enabling quick verification and traceability.
Accepts ultrasound videos, single images, and folders, automatically extracts frames from video inputs and normalizes them into a consistent processing format.
Clinicians can review ultrasound frames with follicles highlighted and labeled, making it easier to confirm findings without stopping the video repeatedly.
Follicles can be corrected when needed (approve/reject, adjust selection), so doctors stay in control of the final clinical output.
For each ultrasound study, the system generates a clear summary of findings (counts, key measurements per ovary/frame range), reducing manual note-taking.
Saved studies can be revisited to compare changes over time during treatment cycles, supporting consistent monitoring and communication with patients.
